LaunchDarkly MCP Server

LaunchDarkly publishes a local MCP server for customers who cannot use its hosted one — specifically its federal and European Union environments, where data residency or compliance rules rule out a vendor-hosted endpoint. Functionally it exposes the same feature-management surface, running on your own machine under your own credentials.

What it exposes. FeatureFlags — get flag status, list, create, patch and delete, which covers the full lifecycle an agent needs to roll a change out behind a flag rather than shipping it bare. AiConfigs — list, create, update and delete AI Configs along with their targeting and variations, LaunchDarkly's mechanism for managing model choice, prompts and parameters as configuration instead of as deployed code. Environments — list environments by project. AuditLog — list audit entries, so an agent can answer who changed which flag and when. CodeReferences — list repositories where flags are referenced, which is how you find the flags that are safe to clean up.

The AI Configs tooling is the part that has become interesting recently: it lets an agent adjust which model or prompt a feature is using, and to whom, without a deploy — the same targeting machinery teams already use for feature rollout, applied to model configuration.

Setup. Runs through npx from the @launchdarkly/mcp-server package, authenticated with an API key from LaunchDarkly's Authorization page (api- followed by a UUID). The key can be passed on the command line or, better, injected from an environment variable so it never lands in a checked-in config file. MIT-licensed.

Note: if you are on standard LaunchDarkly rather than the federal or EU environments, the hosted MCP server is the supported path and this local server is the fallback.